← Projects

LocalProp

Reference implementations of biologically plausible credit-assignment rules for deep networks.

• GitHub
• NeurIPS paper

LocalProp collects clean, benchmarked implementations of learning rules that avoid the biologically implausible weight transport of backpropagation — feedback alignment, target propagation, predictive coding, and our own local error signals from the NeurIPS 2025 paper.

Goals

• One consistent training loop so rules can be compared fairly.
• Benchmarks on vision and sequence tasks with logged compute budgets.
• A teaching resource: each rule has an annotated, minimal notebook.

If you are exploring how the brain might approximate gradient descent, this is a good place to start experimenting.